Impact of segmentation methods on pathological grade prediction in pancreatic ductal adenocarcinoma based on 18F-FDG PET/CT radiomics
10.3760/cma.j.cn321828-20210511-00160
- VernacularTitle:勾画方法对 18F-FDG PET/CT影像组学预测胰腺导管腺癌病理分化程度的影响
- Author:
Zhixin HAO
1
;
Lei LIU
;
Haiqun XING
;
Wenjia ZHU
;
Hui ZHANG
;
Li HUO
Author Information
1. 中国医学科学院、北京协和医学院北京协和医院核医学科、核医学分子靶向诊疗北京市重点实验室 100730
- Keywords:
Carcinoma, pancreatic ductal;
Image processing, computer-assisted;
Positron-emission tomography;
Tomography, X-ray computed;
Deoxyglucose;
Forecasting
- From:
Chinese Journal of Nuclear Medicine and Molecular Imaging
2021;41(8):454-459
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the segmentation methods of pancreatic ductal adenocarcinoma (PDAC) tumor regions in 18F-fluorodeoxyglucose (FDG) PET/CT images, as well as their impact on radiomic features-based pathological grade prediction. Methods:A total of 72 patients (46 males, 26 females, age range: 25-87 years) with pathologically confirmed PDAC and a preoperative 18F-FDG PET/CT scan in Peking Union Medical College Hospital between September 2010 and January 2016 were enrolled retrospectively. The cohort of patients was classified as well differentiated group and non-well differentiated group based on the pathological grade of PDAC, and patients were divided into training set and validation set in the ratio of 3∶1 randomly. Two physicians performed manual contours in the tumor region (referred as region of interest (ROI)_M1 and ROI_M2) and semi-automatic ROIs based on standardized uptake value (SUV) gradient edge search (referred as ROI_G) and 40% threshold applied to the maximum SUV (SUV max; referred as ROI_S) were drawn. The four types of segmentation results were compared in terms of volume and Dice similarity coefficient (DSC). Shape, first-order, and texture features were extracted from PET/CT original and preprocessed images, and the interclass correlation coefficient (ICC) was used to assess each feature′s consistency across all segmentations. Kruskal-Wallis rank sum test, independent-sample t test or z test were used to analyze the data. The area under the receiver operating characteristic curve was used to assess model accuracy, and cross validation was used to assess generalization ability. Results:There were 55 patients in the training set (14 well differentiated cases and 41 non-well differentiated cases) and 17 patients in the validation set (4 well differentiated cases and 13 non-well differentiated cases). A total of 44 selected features were predictive of the pathological grade of PDAC among 20 feature groups. There was significant difference among the volumes of ROI_M1, ROI_M2, ROI_G and ROI_S (10.29(4.01, 19.43), 9.34(4.26, 17.27), 11.86(5.52, 19.74) and 15.08(9.62, 27.44) cm 3; H=18.641, P<0.05). The degree of contour coincidence and feature consistency between ROI_M1 and ROI_M2 were both higher (DSC=0.86 (0.76, 0.90), ICC=0.86 (0.74, 0.94)). Compared to manual contours, the degree of contour coincidence and feature consistency of ROI_G (DSC: 0.86(0.75, 0.91), 0.91(0.85, 0.96); ICC: 0.87(0.72, 0.94), 0.94(0.88, 0.98)) were better. There was no statistically significant difference in model accuracy or generalization ability between ROI_M1 and ROI_G ( z=1.052, t=0.712, both P>0.05). The accuracy of ROI_M2 was better than ROI_G ( z=3.031, P=0.002), but the generalization ability of ROI_M2 was insufficient ( t=3.086, P=0.012). Conclusions:Although the manual contour prediction models are highly accurate, their performance are unstable. Semi-automatic contouring based on gradient can achieve comparable accuracy to manual contouring, and the model′s generalization ability is stronger.
